Sequential Thinking

Sequential Thinking

The core of this skill is not "write more thoughts." It is to let AI continuously advance, allow correction, and finally converge on conclusions in complex problems. The CLI is only the execution carrier; the skill itself defines when to enter this thinking mode and how to avoid degrading sequential thinking into loose output.

Mission

Use this skill to turn complex problems into a bounded, correctable, reviewable reasoning process:

• Clarify the problem first, instead of rushing to answers
• Allow correction and judgment updates during progress
• Compare alternatives when complexity increases, instead of forcing one path
• Converge to conclusions and recommendations within limited steps
• Preserve a replayable reasoning trace at the end

This skill does not solve "can't think." It solves "thinking too scattered, concluding too early, too hard to review."

Core Capabilities

• Iterative advancement: break complex problems into consecutive steps
• Dynamic correction: revisit and revise earlier judgments when new evidence appears
• Branch comparison: compare alternatives before converging
• Context retention: keep clear boundaries and goals in multi-step reasoning
• Conclusion closure: must end with a judgment, not endless divergence

When to Use

Core Decision Rule

Model capability sets baseline; task complexity decides escalation.

Mnemonic

"No CoT -> must use ST Has CoT -> use ST only for complex tasks"

Scenarios where CLI is mandatory

Scenarios where natural CoT is enough (with CoT)

Decision tree

flowchart TD
    A[Need multi-step reasoning?] -->|No| Z[Answer directly]
    A -->|Yes| B{Model has CoT?}
    B -->|No| CLI[Must call ST CLI]
    B -->|Yes| C{Task complex?}
    C -->|Simple: <5 steps, no revision| CoT[Natural CoT]
    C -->|Complex: revision/comparison/replay| CLI

Not suitable for

• Simple fact lookup
• Tasks solvable in one step
• Problems with already obvious path and no multi-step derivation needed
• Pure brainstorming without required convergence

Working Philosophy

• Find the core problem before answers: do not mistake symptom description for root cause diagnosis
• Allow corrections instead of defending bad premises: if earlier steps are wrong, go back and fix
• Reduce complexity before adding solutions: identify the principal contradiction first
• Advance one step at a time: current step states only current judgment
• Must land on conclusions: "I can keep thinking" is not a default exit

Installation & Runtime Model

This skill provides thinking method and invocation constraints for agents; CLI is distributed via npm.

Before use, ensure CLI is installed locally:

npm install -g sequential-thinking-cli

# or
pnpm add -g sequential-thinking-cli

After installation, use sthink as command entry.

CLI Contract

This skill no longer requires AI to hand-write thought JSON. Execution uses main CLI actions:

• start
• step
• replay

start

Accepts only four inputs:

• name
• goal
• mode
• totalSteps

Constraints:

• mode only allows explore, branch, audit
• totalSteps only allows 5 or 8

If unsure which mode to use, default to explore. Use branch only when clearly comparing candidate paths; use audit only when clearly reviewing existing judgments.

step

Accepts only:

• content

All other context should be automatically restored and injected by runtime.

replay

Used to read completed sessions and generate replay docs; optionally export to current directory.

Recommended Workflow

1. Decide whether the problem truly needs sequential-thinking; do not apply by default.
2. If needed, install or confirm local npm CLI.
3. Use `sthink start` with `name`, `goal`, `mode`, `totalSteps`.
4. Use `sthink step` to advance step by step; each step contains only current progress.
5. When new evidence appears, correct earlier judgments.
6. At convergence, output conclusion, risks, and next-step recommendations.
7. Optionally run `sthink replay` to generate/export replay docs.

Examples

The following examples are not for hand-writing JSON; they show where this skill provides real value: advance, correct, converge.

Example 1: Basic reasoning

sthink start --name "query-diagnosis" --goal "locate root cause of query performance degradation" --mode explore --totalSteps 5
sthink step --sessionPath "<session-path>" --content "Do not rush to pick optimization means. First split layers: single SQL degradation, interface-level N+1, or upper-layer amplification. If root cause is unclear, cache/index/rewrite may all become patches."
sthink step --sessionPath "<session-path>" --content "Logs show user-detail endpoint triggers many repeated reads in one request, with clear N+1 signals. But cannot conclude yet; repeated queries may be symptoms. Need to verify whether slowness is from query count or one inherently slow query. Increase total steps."
sthink step --sessionPath "<session-path>" --content "Converged: primary cause is N+1 during list-page batch loading; secondary cause is missing index on related fields amplifying single-query cost. Sequence should remove N+1 first, then add index and validate tail latency."

Example 2: Premise correction

sthink step --sessionPath "<session-path>" --content "After reviewing profiling results, prior judgment needs correction: the real bottleneck is missing index on join columns, which amplifies full-scan costs per join. N+1 still exists but is not first-order bottleneck; priority should move down."

Example 3: Complex change decomposition

sthink start --name "change-impact-analysis" --goal "decompose impact and priorities of complex changes" --mode explore --totalSteps 5
sthink step --sessionPath "<session-path>" --content "User proposed multiple rule changes at once; do not treat them as one type. Split by type: mechanism principles, numeric balancing, interface semantic changes, and doc-implementation drift."
sthink step --sessionPath "<session-path>" --content "Build an impact matrix. Mechanism changes flow into ADR/System Design; numeric balancing affects rule tables/config/test baselines; interface semantic changes are most dangerous because they silently break caller assumptions."
sthink step --sessionPath "<session-path>" --content "Converged: first handle items changing system boundaries or call semantics, then numeric/experience items. Sequence should fix docs/contracts before balancing, otherwise implementation and review build on drifting premises."

Example 4: Branch comparison

sthink start --name "performance-tradeoff" --goal "compare priority of cache stopgap vs query optimization" --mode branch --totalSteps 5
sthink step --sessionPath "<session-path>" --content "Option A: introduce cache first to reduce peaks. Fast effect, low interface intrusion, good for stopgap; downside is complexity in consistency/invalidation and may preserve accidental complexity. Option B: optimize indexes and rewrite queries directly. Root-cause oriented and cleaner long-term, but requires careful validation of write amplification, lock contention, and regression risk."

Storage & Export Boundary

• runtime automatically saves session state and step records
• replay docs can be generated after completion
• replay supports export to current directory for review/reuse

Heuristic Reminders

These are heuristic prompts, not hard constraints. The key is reducing wasted loops and approaching conclusions.

• Problem-definition prompt: Are you describing symptoms or locating root cause?
• Evidence prompt: Is current judgment based on facts/observations or guesses?
• Boundary prompt: Is impact local module, single system, or cross-system?
• Complexity prompt: Are you reducing essential complexity or adding accidental complexity?
• Convergence prompt: Is there enough to conclude, or still ineffective divergence?

Tips

• Do not hand-write thought JSON; let CLI runtime handle pacing, persistence, replay
• Do not treat sequential-thinking as default; use only when multi-step convergence is truly needed
• If unsure mode, start with explore
• step content should only express current progress, not repeated full context
• If a premise is wrong, state correction explicitly
• At convergence, clearly output conclusions, risks, next actions
• replay is available only for completed sessions

Source: Haaaiawd/ANWS — distributed by TomeVault.